Sparse light field imaging often limits the quality of 3D scene reconstruction due to insufficient viewpoint coverage, resulting in incomplete or inaccurate reconstructions. This work introduces a hybrid CNN–LSTM-based framework to address this issue by generating novel camera poses and the corresponding synthesized novel views, effectively densifying the light field representation. The CNN extracts spatial features from the sparse input views, while the LSTM predicts temporal and positional dependencies, enabling smooth interpolation of novel poses and views. The proposed method integrates these synthesized views with the original sparse dataset to produce a comprehensive set of images. Our approach was evaluated on several datasets, including challenging datasets. The inference capability of our method was tested extensively, and it showed good generalization across diverse datasets. The effectiveness of the framework was evaluated not only with local light field fusion (LLFF) but also with NeRF and 3D Gaussian Splatting, which are considered state-of-the-art reconstruction methods. Overall, the enriched dataset generated by our method led to consistent improvements in 3D reconstruction quality, including higher depth estimation accuracy, reduced artifacts, and enhanced structural consistency. Most importantly, LSTM-based approaches have so far attracted limited attention in the context of generating novel views. While LSTMs have been widely applied in sequential data domains such as natural language processing, their use for image generation conditioned on camera poses remains largely unexplored, which underscores the novelty and significance of the proposed work. This approach provides a scalable and generalizable solution to the sparsity problem in light fields, advancing the capabilities of computational imaging, photorealistic rendering, and immersive 3D scene reconstruction. The results firmly establish the proposed method as a robust and versatile tool for improving reconstruction quality in sparse-view settings.

Gaussian Splatting (GS) has emerged as an efficient and compelling method for novel view synthesis, offering real-time rendering and impressive visual fidelity. However, original GS struggles with accurately capturing the geometry of large, flat, or textureless surfaces, often producing floating artifacts, blurriness, and inconsistent structures. To overcome this fundamental limitation, we introduce AnchorGS, a framework that anchors the optimization process to explicit geometric plane priors. AnchorGS operates in a multi-phase hierarchical manner, progressively refining Gaussians using a suite of context-aware regularization terms: plane attraction, normal alignment, and explicit thinness constraints. Furthermore, it incorporates a specialized densification strategy to systematically fill holes and refine the Gaussian distribution on identified planar surfaces. Experiments demonstrate that AnchorGS significantly enhances reconstruction quality, yielding sharper and more geometrically accurate planar surfaces while maintaining real-time rendering performance. Our method effectively mitigates common artifacts in scenes containing prominent planar structures, paving the way for more geometrically robust scene representations.

The robust estimation of dense correspondences between images is a cornerstone of computer vision, enabling applications from 3D reconstruction to autonomous navigation. While learning-based methods have recently shown promise, they often require extensive training data and specialized hardware. Conversely, classical methods such as SIFT and ORB, though robust to large viewpoint changes, can be computationally expensive and may not produce the match density required for high-fidelity surface reconstruction. In this paper, we propose KRAFT (Keypoint Robust and Adaptive Feature Tracking), a novel, computationally efficient feature matching algorithm designed specifically for small-baseline stereo scenarios, such as those encountered in stereo vision systems and video sequences. The algorithm first generates a dense field of candidate matches by describing salient, gradient-based keypoints with their raw pixel patch neighborhoods, a lean strategy designed to maximize match quantity. To ensure exceptional precision, KRAFT then introduces its sophisticated optional filtering cascade, KRAFT-Clean. This cascade performs a powerful two-step verification on the candidate matches, first leveraging on-demand depth information to enforce geometric consistency in 3D space. This is followed by a final validation using a unique quadrant-based NCC that assesses a patch’s internal structural integrity, a step that makes the algorithm exceptionally robust to the partial occlusions and non-rigid deformations common in real-world scenes. We perform a comprehensive evaluation on diverse datasets, including KITTI, Middlebury, and a custom human-centric dataset. Our results demonstrate that KRAFT consistently outperforms standard detectors in precision across all scenarios tested. Furthermore, in its target domain of human-centric scenes, it also produces a significantly higher quantity of correct matches, achieving a state-of-the-art mean precision of 98.9%. KRAFT represents a powerful and efficient alternative for applications requiring dense, high-quality feature matching. It serves as an ideal front-end for geometrically grounded novel view synthesis, with the potential to significantly improve the underlying Structure-from-Motion results that are critical for state-of-the-art techniques like Neural Radiance Fields (NeRF) and 3D Gaussian Splatting.